WebRobot

Learn

No magic: how things work, one step at a time

Interactive lessons on what sits behind the words: how a network learns, how an app is designed. Each one is small enough to follow every piece on screen, and computes everything live in your browser.

You need no installation. Each path builds on its own lessons, in order.

Transformers: a beautiful invention, not magic

Behind a language model there are multiplications, a rule for sharing out the error, and a lot of repetition. Here we build it one piece at a time, on networks small enough to follow every number on screen.

Each lesson is one interactive 3D network that you run step by step, with no maths beyond multiplying: the derivatives are explained where they appear.

Lesson 1

Backpropagation

A 1-2-1 network: how the loss travels back, what reaches the middle layer, how each gradient is computed.

Open the lesson →

Lesson 2 · Coming next

Attention

We add attention to the same small network: queries, keys and values are just more weights, and the gradient flows through them the same way.

Lesson 3 · Coming next

KV cache, and the Transformer

A small KV cache: when generating one token at a time, keys and values already computed are kept instead of recomputed. Then the pieces come together into a Transformer.

Designing an app with a manifest

How an app is described instead of programmed. One small app, built one entity at a time on a graph that grows, checked live by the platform’s own validator.

Lesson 1

The manifest, step by step

People, data, pages, agents, events, approvals, billing, the boundary: nine questions, nine steps, a graph that grows, and the typical mistake of each step.

Open the lesson →

Lesson 2

Data: whose is each record

Collections and fields; ownedBy user, shared, app, org; access rules and where. A matrix shows who reads and writes each record, recomputed with the runtime’s own rule.

Open the lesson →

Lesson 3

Pages: views over the data

The same records as a table, cards, KPIs, bars or a chart, with formats, filters, groups and a teaser for the paying plan; landing pages made of blocks.

Open the lesson →

Lesson 4

Appearance and languages

Palette, fonts, buttons, a dark theme, the header, and texts in every language — styled live by the runtime’s own theme function.

Open the lesson →

Lesson 5

Agents: task, tools, result

goalTemplate and its placeholders, the app tools, structured output and writesTo; teams, delegation, memory and Python slots.

Open the lesson →

Lesson 6

Events and checkpoints

Triggers, schedules and personal agents from templates; approvals, and how a decision resumes the work.

Open the lesson →

Lesson 7

Money: billing

Free, credits, subscriptions and per-transaction fees through Stripe Connect; metering on real usage; limits per plan.

Open the lesson →

Lesson 8

The boundary

External hosts and keys, ETL jobs on the platform, phone, Telegram and email channels, the phone app.

Open the lesson →

Lesson 9

From words to a manifest

What the builder does with a description in plain words, how its manifest passes the three checks, and how a failed check is repaired.

Open the lesson →

Lesson 10

Summary: what you can build

The nine axes of the model, the families of apps it covers, where it stops — and what the builder produced from 370 real startups.

Open the lesson →